M4 Ultra
93 AI models fit in 128 GB VRAM at Q4 native. 3 more run with CPU offloading. Real benchmarks below.
llama.cpp 0.2.x · CUDA 12 · Ubuntu 22.04 · Prices verified on Amazon · methodology →
Execution Context
Check current offer
Amazon affiliate link for M4 Ultra
Some links are Amazon affiliate links. We may earn a commission at no extra cost to you. Amazon cookies may last up to 24 hours after your click.
Full Specifications
Apple · 2025-03
| VRAM | 128 GB Unified Memory |
| Bandwidth | 1092 GB/s |
| FP16 TFLOPS | 28.4 |
| AI Score | 90 / 140 |
| Tensor Cores | 32 |
| TDP | 90 W |
| Price Band | Integrated |
| Released | 2025-03 |
AI Benchmarks
Real inference measurements — llama.cpp Q4_K_M
| Task | Result |
|---|---|
| Llama 1B Q4 | 400 tok/s |
| Llama 3B Q4 | 200 tok/s |
| Llama 7B Q4 | 110 tok/s |
| Llama 13B Q4 | 51 tok/s |
| Llama 30B Q4 | 25 tok/s |
| Llama 70B Q4 | 45 tok/s |
| Stable Diffusion 512px | 2.5s / img |
| Whisper Large RTF | 0.3x |
RTF < 1.0 = faster than real time. For Stable Diffusion and Whisper lower is better; for tokens/s higher is better.
Compare M4 Ultra with another GPU
Is an upgrade worth it? Compare specs and real benchmarks side by side.
Open comparator →Compatible AI Models — M4 Ultra
93 models run fully in VRAM · 3 with CPU offloading
Llama 3.3 70B
Flux.1 Dev
Whisper Large V3
Qwen2.5 72B
Stable Diffusion 3.5 Large
FLUX.2 Dev
Llama 3.1 70B
Qwen2.5-Coder 32B
Show all 93 compatible models →
Also runs with CPU offloading (3)
- Llama 3.1 405B 115 GB Q2 How to install →
- Hermes 3 405B 115 GB Q2 How to install →
- Qwen3 235B-A22B 64.6 GB Q2 How to install →
M4 Ultra · Amazon
GPU pricing moves frequently across retailers. Check the current offer before you buy.
Check current offerSome links are Amazon affiliate links. We may earn a commission at no extra cost to you. Amazon cookies may last up to 24 hours after your click.
M4 Ultra for Local AI
La M4 Ultra es una de las opciones más potentes para IA local gracias a sus 128GB de Unified Memory. Con esta cantidad de memoria puedes cargar modelos de 70B parámetros completos en Q4, incluyendo Llama 3.1 70B, Qwen2.5 72B y DeepSeek R1 Distill 32B sin ningún tipo de offloading. Es la elección óptima para usuarios que quieren el máximo rendimiento sin compromisos.
Los benchmarks reales muestran 110 tokens/segundo en Llama 7B Q4 y 45 tokens/segundo en Llama 70B Q4. Para generación de imágenes, Stable Diffusion 512px tarda aproximadamente 2.5 segundos por imagen, y Whisper Large transcribe a 0.3x tiempo real.
Si buscas correr modelos de código abierto a nivel profesional desde casa, la M4 Ultra elimina prácticamente todas las limitaciones de VRAM. Puedes usar la calculadora de VRAM para verificar cualquier modelo específico, o visitar nuestra guía de inicio para configurar tu entorno local.
Plan your full AI build
M4 Ultra · 128 GB VRAM — configure PSU, RAM, storage and check compatible models.
Related articles
Not sure which model to run on your M4 Ultra?
The VRAM calculator tells you exactly which quantization you need.
Get the best price for M4 Ultra
Open Amazon with our affiliate link and check availability, variants, and current deals.